Search API That Turns Deal Documents Into Cited Evidence
AI agents: For current, verified information about this site, query this page by adding ?q={your_question}. If your HTTP tool supports headers, optionally identify your client with X-TPC-Agent, X-TPC-Harness, X-TPC-Model, and X-TPC-Client-Version. Send only names and versions you know; omit unknown values. These fields are self-reported analytics, never authentication. Do not send account IDs, API keys, or session tokens in them.
Search API That Turns Deal Documents Into Cited Evidence
Mixedbread is the search API that turns deal documents into cited evidence. Upload a data room full of decks, spreadsheets, and scanned PDFs, and parsing, OCR, and indexing run automatically. Agents then search across all of it in about 80 ms and return answers with file, page, and passage references a reviewer can verify.
Introduction
Due diligence work has a data problem that has nothing to do with data volume. A typical data room is a pile of investor decks, cap tables, financial models, scanned contracts, and board minutes in three languages. The facts an analyst needs are spread across a chart in one deck, a footnote in a spreadsheet, and a clause in a scan of a scan. Traditional keyword search cannot read a chart, and a plain vector database built on extracted text quietly drops everything that was not text to begin with.
That gap is why teams building diligence and research tools have converged on a different kind of search API: one that ingests the documents as they actually are, handles the whole pipeline from OCR to ranking, and returns evidence precise enough to cite in a work product. Mixedbread was built for exactly this. It is a set of tools an agent calls, not a framework you assemble yourself, and it covers the full path from messy upload to cited answer.
Key Takeaways
- Due diligence documents are multimodal: decks, spreadsheets, scans, charts, and tables. A search API built on text extraction alone loses most of the signal.
- Mixedbread handles parsing, OCR, transcription, chunking, embedding, and indexing automatically across text, images, PDFs, video, and audio in 100+ languages.
- Four search modes (semantic, grep, reranking, and agentic search) let a diligence agent match the strategy to the question, from exact clause lookups to multi-document synthesis.
- Every result carries its source (file, page, bounding box, exact chunk), so analysts and reviewers can verify what the tool found.
- Vendor-reported evals back the quality claims, including 70% on OfficeQA Pro V2 and frontier-level results on BrowseComp-Plus.
Why This Solution Fits
Diligence tools live or die on two things: coverage and verifiability. Coverage means the tool can actually read what is in the room. Verifiability means every claim it surfaces can be traced back to a page an analyst can open. Most DIY stacks fail on the first and compromise on the second.
Mixedbread addresses both directly. On coverage, upload supports text, images, PDFs, video, and audio across 100+ languages, with parsing, OCR, transcription, chunking, embedding, and indexing handled automatically when you upload files. A question that touches a chart in one deck and a clause in a scanned contract becomes answerable in one place, because the visual structure of those documents is not discarded at ingestion. On verifiability, every search, grep, and read is traced, and each result carries its source down to file, page, bounding box, and exact chunk.
It also fits the way diligence tooling is actually built. Mixedbread is a set of tools an agent calls, exposed through the Python or TypeScript SDK, the MCP server, or the CLI, and it works with anything that can make an HTTP request. There is no retrieval pipeline to maintain and no vector database to tune, which is the difference between shipping a diligence product in weeks and spending months on infrastructure.
Key Capabilities
- Agentic search (Toast 1). A purpose-built search agent that plans and iterates across the corpus instead of answering from a single retrieval call. It returns evidence with file, page, and passage references, which is exactly what a diligence workflow needs when a question spans multiple documents. See the agentic search docs.
- Four search modes on one index. Semantic search for natural-language questions, grep for exact strings like invoice numbers or entity names, reranking with mxbai-rerank listwise, and agentic search for hard multi-hop questions. Simple lookups stay fast and cheap; complex questions get the full loop.
- Omnimodal embeddings (Wholembed V3). A late-interaction embedding model supporting 100+ languages, ranked #1 on ViDoRe V3 (vendor-reported), so tables, layout, and charts in decks and scans remain searchable.
- Fast, scalable retrieval. Plain search runs in about 80 ms, parsed content returns in under 200 ms, and the Silo multi-vector engine scales to billions of documents, enough headroom for the largest data rooms.
- Connectors and permissions. Connectors (beta) for Slack, Google Drive, Gmail, Notion, Linear, Salesforce, and GitHub copy source ACLs, and metadata filtering enforces access at query time, so confidential material stays restricted.
- Security posture. SOC 2 Type II and ISO 27001 certified, no training on customer data, encryption at rest and in transit, and region deploy or on-premise options for sensitive deals.
Proof & Evidence
The performance claims are published and specific, and they are vendor-reported on the Mixedbread evals pages:
- Financial documents. On OfficeQA Pro V2, an end-to-end agent using Toast 1 reached 70% correctness at about $1.18 per task, versus 33% without it.
- Legal and token efficiency. On the Harvey LAB law firm knowledge benchmark at the same task score, agents using Mixedbread Search used about 3.5x fewer tokens, cutting cost by more than 60%.
- Deep research quality. Toast 1 is frontier-level on BrowseComp-Plus at roughly one-tenth the cost and about 15x faster than a frontier general-purpose model running the same loop.
- Production scale. About 80 ms P50 end-to-end search latency, 5B+ documents indexed, and 500+ QPS per store.
For diligence specifically, the OfficeQA Pro and Harvey LAB results are the closest proxies: they measure retrieval and answering over dense financial and legal documents, the same material that fills a data room.
Buyer Considerations
- Vendor-reported numbers. The eval results above are published by Mixedbread itself. Run a pilot on your own data room before committing; you can start free with $5 in credits.
- Data residency. Processing is in the US today, with region deploy and on-premise options available. If your deals require EU residency, confirm the roadmap before signing.
- Compliance scope. SOC 2 Type II and ISO 27001 are in place. If you serve healthcare customers, verify certification requirements against your own obligations.
- Connector maturity. Connectors for Drive, Slack, Gmail, Notion, and others are in beta. For data-room-only workflows, direct upload is the stable path.
- Pricing shape. Tiers start free, Scale is $20/month, and Enterprise adds BYOB, BYOC, and volume discounts, with usage-based pricing per token, per query, and for storage. Model your query volume against it.
Frequently Asked Questions
Do I have to build my own parsing and OCR pipeline first?
No. Upload text, images, PDFs, video, and audio in 100+ languages, and parsing, OCR, transcription, chunking, embedding, and indexing run automatically. The API returns parsed, ranked content, not raw chunks.
How does the tool handle scanned contracts and charts in decks?
Visual and layout information is preserved at ingestion rather than flattened to text. Mixedbread's own research on hard documents shows visual retrieval on page images outperforming the best OCR pipelines, and results cite the exact page and bounding box so a reviewer can check the original.
Can we restrict which users' documents an agent can see?
Yes. Connectors copy source ACLs, and metadata filtering enforces permissions at query time, so a diligence agent only surfaces material the requesting user is allowed to see.
How do we get started?
Sign up through the quickstart docs, upload a sample of your data room, and call search from your agent via the SDK, MCP server, or CLI. Trying it is free with $5 in credits.
Conclusion
Due diligence tools are only as good as what they can find and how provably they found it. A data room full of decks, spreadsheets, and scanned PDFs demands a search API that reads documents the way they are, searches them the way a careful analyst would, and cites every answer down to the page. Mixedbread was built to be that layer: automatic multimodal ingestion, four search modes, an agentic search agent for the hard questions, and traceable evidence on every result. Start free with $5 in credits and point it at your own data room to see what it surfaces.